Ant Colony-Based Hyperparameter Optimisation in Total Variation Reconstruction in X-ray Computed Tomography

Manasavee Lohvithee, Wenjuan Sun, Stéphane Chrétien, Manuchehr SoleimaniSensors 2021 · DOI

TL;DR

This paper introduces an automated method using Ant Colony Optimization (ACO) to find optimal hyperparameters for Total Variation (TV) based X-ray Computed Tomography (XCT) reconstruction. This approach significantly speeds up the process compared to traditional methods like cross-validation, while achieving comparable image quality, making it more practical for real-world applications.

What Problem Does This Paper Address?

X-ray Computed Tomography (XCT) is a powerful imaging tool, but reducing the radiation dose to patients is a major concern. This often leads to using limited projection data, which can result in noisy or artifact-ridden images when using standard reconstruction methods. Iterative reconstruction methods, particularly those employing Total Variation (TV) regularization, offer better image quality with limited data by preserving edges and reducing noise. However, these methods are highly sensitive to their hyperparameters, which control the regularization strength and convergence. Manually tuning these hyperparameters is a tedious, time-consuming, and data-specific process, requiring expert knowledge. This paper addresses the challenge of efficiently and automatically finding optimal hyperparameters for TV-based XCT reconstruction algorithms.

Methods

The core of this paper is the development of a computer-aided hyperparameter tuning method that leverages the Ant Colony Optimization (ACO) algorithm. The authors use the Adaptive-weighted Projection-Controlled Steepest Descent (AwPCSD) algorithm as their TV-based reconstruction method.

Here's a breakdown of the approach:

  • AwPCSD Algorithm: This iterative algorithm has two phases:
  • SART (Simultaneous Algebraic Reconstruction Technique): Enforces data consistency by minimizing the difference between projected images and measured data, while also ensuring non-negativity of the reconstructed image.
  • TV Optimization: Minimizes the adaptive-weighted TV norm of the image using a steepest descent approach to preserve edges.
  • Ant Colony Optimization (ACO): This metaheuristic approach is used to search for the optimal set of hyperparameters.
  • Artificial Ants: A colony of artificial ants explores different hyperparameter configurations.
  • Pheromone Trails: Ants leave "pheromones" on paths representing hyperparameter choices. Higher scores for a configuration lead to stronger pheromone trails, attracting more ants in subsequent generations.
  • Scoring: Each ant runs the AwPCSD algorithm with its chosen hyperparameters, and the reconstructed image is scored based on its correlation coefficient (CC) with a reference image.
  • Iterative Refinement: The algorithm iteratively refines the pheromone trails and probabilities, guiding the ants towards better hyperparameter configurations.
  • Hyperparameter Selection: The AwPCSD algorithm requires five hyperparameters. The authors fix three of them based on prior work (relaxation parameter $\beta$, reduction factor $\beta_{red}$, and scale factor $\delta$). The ACO algorithm is then used to optimize the remaining two: data-inconsistency-tolerance parameter ($\epsilon$) and TV sub-iteration number ($n_g$).

Datasets

The study utilized datasets simulated from the 4D Extended Cardiac-Torso (XCAT) Phantom [16,17].

  • Training: One dataset was used for training the hyperparameter optimization algorithm.
  • Testing: Two additional XCAT phantom datasets, with different parametrizations (male and female phantoms), were used for testing the robustness of the trained hyperparameters.
  • Noise Simulation: Poisson and Gaussian noise were added to the projection data to simulate realistic imaging conditions.

Key Insights and Ideas

  • ACO as an Efficient Search Strategy: The paper demonstrates that ACO can effectively navigate the complex, multi-dimensional hyperparameter space of TV-based reconstruction algorithms, a task that is notoriously difficult and time-consuming with manual tuning.
  • Data-Specificity of Hyperparameters: The results highlight that while a set of hyperparameters optimized for one dataset can be applied to others with similar contexts, fine-tuning is often necessary to achieve the absolute best results for a specific dataset.
  • Trade-off Between Quality and Speed: The proposed ACO method achieves image quality comparable to the gold standard cross-validation method but is significantly faster, making it more practical for clinical applications where time is often a critical factor.
  • Automation Reduces User Burden: The fully automatic nature of the ACO-based tuning process eliminates the need for expert intervention or prior knowledge of the reconstruction algorithm from the user.

Key Results

  • Image Quality: The proposed ACO-based method produced reconstructed images that were visually and quantitatively superior to the Conjugate Gradient Least Square (CGLS) algorithm and the AwPCSD algorithm with an arbitrary hyperparameter setting. The quality was comparable to, and in some cases slightly better than, the cross-validation method.
  • Computational Speed: The ACO approach was over 10 times faster than the cross-validation method. While CGLS was the fastest, it produced significantly noisier images. The ACO training took approximately 1.45 hours, whereas cross-validation took around 47.15 hours.
  • Robustness to Noise: The optimal hyperparameter set found by the ACO method remained robust against increasing levels of noise in the projection data, maintaining superior image quality compared to other methods.
  • Robustness to Angle Arrangements: The trained hyperparameters also performed well across different projection angle arrangements, indicating the generalizability of the optimized settings.
  • Data-Specific Optimization: While the trained hyperparameters were applicable to new phantom datasets, directly applying the ACO method to the new data yielded even better results, underscoring the benefit of data-specific optimization.

Limitations and Future Work

  • Fixed Hyperparameters: The study fixed three hyperparameters ($\beta$, $\beta_{red}$, $\delta$) based on previous work. Different datasets might require different optimal values for these fixed parameters.
  • Limited Scope for Initial Proof of Concept: The proposed method was initially developed to optimize two hyperparameters ($\epsilon$ and $n_g$) to reduce complexity.
  • Future Work Suggestions:
    • Broader Hyperparameter Ranges: Extend the method to include more/broader ranges of hyperparameters for implementation. This would increase complexity and computational time but could potentially yield even better results.
  • Real CT Measurement Data: Evaluate the method on real CT measurement data to assess its performance in more realistic noise conditions.

Summary updated

BibTeX
@article{lohvithee2021ant,
  author = {Manasavee Lohvithee and Wenjuan Sun and Stéphane Chrétien and Manuchehr Soleimani},
  title = {{Ant Colony-Based Hyperparameter Optimisation in Total Variation Reconstruction in X-ray Computed Tomography}},
  journal = {Sensors},
  year = {2021},
  doi = {10.3390/s21020591},
}
Figure